Vercel AI SDK

The AI SDK does not hand you a provider client to wrap. It hands you a language model. So Metergraph attaches as AI SDK middleware rather than through wrap().

This is TypeScript only. There is no Python equivalent, because the AI SDK is a TypeScript library.

Attach the middleware

src/llm.ts

import * as mg from "metergraph";
import { generateText, wrapLanguageModel } from "ai";
import { openai } from "@ai-sdk/openai";

const model = wrapLanguageModel({
  model: openai("gpt-5-mini"),
  middleware: mg.vercelAISDKMiddleware(),
});

await mg.trace("support-answer", () =>
  generateText({ model, prompt: "Help this customer" }));

generateText and streamText are both covered: the middleware implements wrapGenerate and wrapStream, so the two calls take the same capture path.

The middleware initializes Metergraph itself the first time it is created, so you can hand it init() options directly instead of calling init() separately:

mg.vercelAISDKMiddleware({
  repository: "owner/repository",
  environment: "production",
});

If your application already calls mg.init() centrally, call it first and pass the middleware nothing.

AI SDK versions

The AI SDK’s middleware protocol has its own version number, separate from the AI SDK release. The middleware defaults to protocol v3, which is what AI SDK 6 uses and what AI SDK 7 still accepts. AI SDK 5 uses protocol v2, so it is the one case where you have to say something.

AI SDK	Call	Node.js
5	vercelAISDKMiddleware({ aiSdkVersion: 5 })	18 and newer
6	vercelAISDKMiddleware()	18 and newer
7	vercelAISDKMiddleware()	22 and newer

Metergraph itself supports Node.js 18 and newer. The Node 22 floor in the last row is the AI SDK 7 release’s own requirement, not Metergraph’s.

aiSdkVersion: 5 is the only accepted value for that option. It exists to map AI SDK 5 onto protocol v2 without asking you to know the protocol numbers.

The advanced escape hatch

specificationVersion sets the raw middleware protocol version directly, and accepts "v2", "v3" or "v4". Reach for it only if a future AI SDK needs a protocol the table above does not cover.

mg.vercelAISDKMiddleware({ specificationVersion: "v4" });

The two options cannot be combined

Passing both aiSdkVersion and specificationVersion throws a TypeError immediately, before any model call. Pick one.

Applying it once, not per call site

Repeating wrapLanguageModel at every call site is how instrumentation gets missed. Two patterns avoid it.

A provider registry takes the middleware once, and every model it hands out is captured:

src/models.ts

import { createProviderRegistry, gateway } from "ai";
import { anthropic } from "@ai-sdk/anthropic";
import { openai } from "@ai-sdk/openai";
import * as mg from "metergraph";

const registry = createProviderRegistry(
  { anthropic, gateway, openai },
  { languageModelMiddleware: mg.vercelAISDKMiddleware() },
);

export const model = registry.languageModel("gateway:anthropic/claude-sonnet-4.5");

An existing model factory is the other place to put it: wrap the single return path your application already funnels model construction through. Use the registry for new multi-provider code. Do not invent a factory purely for Metergraph.

Both patterns have a runnable example: registry, existing factory.

What a row records

The middleware records the model ID from the language model, the standardized prompt and the call settings.

It deliberately does not record providerOptions, headers or abortSignal. Those carry credentials, callbacks and values that do not serialize, and the standardized prompt and settings are enough to reconstruct the call.

The endpoint is recorded as ai.doGenerate or ai.doStream.

Provider naming

The AI SDK’s provider string is normalized so rows line up with directly wrapped clients:

• amazon-bedrock, aws-bedrock and aws all become bedrock
• gemini, google-genai and anything starting google. become google
• a prefixed name such as openai.chat becomes openai
• a gateway provider (gateway, vercel, vercel-ai-gateway) resolves through the model ID’s creator prefix, so anthropic/claude-sonnet-4.5 becomes anthropic

See Vercel AI Gateway.

Multi-step tool loops

Each provider request in a multi-step tool loop becomes its own row. That is the point: a loop that quietly runs six model calls shows six costed rows, not one. Wrap the whole operation in mg.trace() so they read as one workflow, and mg.track() or mg.route() to name it.

Streaming

The middleware wraps the returned stream rather than consuming it. Your for await loop sees every part, unchanged and in order. Time to first token comes from the first part that carries real output: a text or reasoning delta, a tool call, a tool result, a file or a source. The row is finalized when the stream ends, and a cancelled stream is finalized with an abandoned status rather than being dropped.

Do not double-instrument

Middleware or wrap(), never both

If a model goes through the middleware, do not also call mg.wrap() on the underlying provider client. One provider request would produce two rows and double its cost in every total.

Fail-open

The middleware is telemetry. If capture cannot start, the call runs uninstrumented. If capture cannot finish, the row is dropped. Neither reaches your generateText result or your stream, and a fault while inspecting one stream part yields that part unchanged and moves on.

The middleware has no runtime dependency on ai. It is a plain object matching the middleware shape, so installing Metergraph does not pull the AI SDK in.

Next

• Vercel AI Gateway for gateway-routed models
• Name a route for track() and route()
• Serverless and short-lived jobs for Vercel functions and edge runtimes
• TypeScript SDK for the full signature